Deep Learning to improve Experimental Sensitivity and Generative Models for Monte Carlo simulations for searching for New Physics in LHC experiments

¹ Instituto de Física Corpuscular (IFIC) - Universitat de València (UV) and CSIC, Valencia, Spain
² Universitat Politecnica de Valencia (UPV), Valencia, Spain
³ Universidad de Alcalá de Henares (UAH), Spain
⁴ Universitat de València (UV), Spain

^* e-mail: jose.salt@ific.uv.es

Published online: 6 May 2024

Abstract

ML/DL techniques have shown their power in the improvement of several studies and tasks in HEP, especially in physics analysis. Our approach has been to take a number of the ML/DL tools provided by several open-source platforms and apply them to several classification problems, for instance, to the tt¯ resonance extraction in the LHC experiments. Gradient-boosting Trees, Random Forest, Artificial Neural Networks (ANN), etc. have been used and optimized by means of adjusting several hyperparameters to control overfitting. On top of this, data simulation with traditional models is computationally very demanding, making the use of generative models an alternative for generating simulated Monte Carlo events with similar quality at a lower computational cost. This could help to produce more simulated data statistics available for better sensitivity and more accurate assessment of systematic errors in potential Physics Beyond Standard Model discoveries. In this work, we study the use of generative models based on Deep Learning as faster Monte Carlo event generators in the LHC context, reducing the time and energy cost of currently used methods. In particular, we focus on different configurations of Variational Autoencoders, taking as a starting point the well-known β-VAE and proposing the α-VAE as a new and simpler VAE architecture that improves the results in some experiments. Considerations will be made about the reliability of these simulated data when they are produced with very high statistics.